Fig. 2

Automated image analysis for five of the most prevalent GPIBDs. A model for the classification of the gene–phenotypes was repeatedly trained and cross-validated on patient subsets that were randomly down-sampled to the same cohort size of n = 10. A mean accuracy of 0.44 was achieved, which is significantly better than random (0.20). For explanatory purposes, the rows of the confusion matrix start with instances of previously published or newly identified individuals with GPIBDs. If the predicted gene matches the molecularly confirmed diagnosis, such a test case would contribute to the true positive rate, shown on the diagonal. Actual affected individual photographs were used to generate an averaged and de-identified composite photo and are shown at the top of the columns. The performance of computer-assisted image classification is significantly better than expected under the null model of perfect heterogeneity and indicates a gene-specific phenotypic substructure for the molecular pathway disease. Higher false positive error rates occur between genes of the same phenotypic series, HPMRS and MCAHS, as indicated by the dendrogram